Improvement And Application Of Particle Swarm Algorithm Learning Strategy

Particle swarm optimization algorithm(PSO)is a group intelligent algorithm that simulates the behavior of bird groups.It has the advantages of less parameters,simple search mechanism and strong global search ability.So it has been widely concerned and studied in recent years.It has been already successfully applied to many fields,such as function optimization,image processing,path optimization and so on.However,the PSO algorithm lacks a solid theoretical basis.Therefore,it is very difficult to balance the relationship between convergence accuracy and population diversity of PSO during the optimization process,.Meanwhile,there is also a common phenomenon of PSO that the population is easily trapped in a local optimal solution or premature.On the basis of last researches,this paper makes a series of improvements for PSO in continuous optimization problems and discrete optimization problems.The main work is as follows:(1)A fitness-based multi-role particle swarm optimization(FMPSO)is proposed.In FMPSO,a new component,named as “subsocial-learning” part,is added to particles’ velocity update aiming to fully utilize the information of the dominant paticles.Moreover,the role assignment mechanism is introduced to this algorithm,in which three roles,i.e.,leader,rambler,and follower,are assigned for particles at each generation based on their fitness.The aim of this strategy is to help population carry out various search mechanisms.During the evolutionary process,two tuning operators are introduced to adjust particles’ roles and objective dimensions.Moreover,a local searching operator based on BFGS Quasi-Newtom method is proposed to refine the best solution at the later evolutionary stage.Experimental results show that FMPSO outperforms other PSOs on a majority of functions in CEC2005 and CEC2013 testsuites,in terms of the global search ability,solution accuracy,and convergencespeed.(2)A fitness-based covariance matrix adaptation multi-role particle swarm optimization(CMA-FMPSO)is proposed.In CMA-FMPSO,the learning mechanism of CMA-ES is introduced into the FMPSO to futher improve the performance of the algorithm.Firstly,the multivariate normal distribution is used for population initialization and the subsequent population updation.Secondly,the multi-modal individual optimal updating mechanism is used to update Pbest,so that the superior information is used in a wide way.Experimental results show that the CMA-FMPSO dominates other comparison algorithms on a majority of functions in CEC2017 testsuite in terms of the performation in complex problems and convergence speed.(3)A probabilitic and deterministic hybrid particle swarm optimization(PDHPSO)is proposed.Based on the discrete hybrid PSO algorithm with genetic operators,there is a series of related strategies improvement in PDHPSO to solve the traveling salesman problem(TSP).In PDHPSO,probabilistically initialized based on coordinate information is used to get more useful information for the algorithm.Moreover,there are some improvements for crossover operator and mutation operator,respectively.In this way,there will be more effective information to guide population evolution.Experiments show that the improved discrete hybrid PSO algorithm has obvious advantages in solving the TSP.